Gauss What You Need: Compact Gaussian Splatting Across Scene Scales

Afif Boudaoud; Jiayi Liu; Alexandru Calotoiu; Torsten Hoefler

cs.CVSep 25, 2026

Gauss What You Need: Compact Gaussian Splatting Across Scene Scales

Authors: Afif Boudaoud, Jiayi Liu, Alexandru Calotoiu, Torsten Hoefler

Organizations: ETH Zürich

Abstract

3D Gaussian Splatting reconstructs a scene as a collection of Gaussian primitives from a set of posed photographs called the capture. The number of primitives used to represent the scene affects reconstruction quality, storage, and rendering cost. How to select this number automatically across capture scales remains unresolved: configurations effective on standard benchmarks can leave larger captures with too few Gaussians to reconstruct fine details. We observe that the surface to represent, given by the capture's extent and resolution, is known before training, whereas its content complexity becomes apparent during training, through the reconstruction quality on the training views. We introduce TangoGS, which combines capture-derived model sizing with training-based adaptation: the capture determines the scale of the model, and training feedback determines its final size within that scale. Before training, TangoGS derives a learning allowance for model growth from the capture's total pixels after discounting views that re-observe the same scene points. During training, reconstruction quality guides how many Gaussians to add and remove. On 13 standard benchmark scenes, TangoGS matches the mean PSNR of the best-performing evaluated baseline, LeGS, with $48\%$ fewer Gaussians. On eight large captures, the same configuration automatically scales to larger models when necessary, achieving the highest mean PSNR among evaluated methods: $0.54$ dB above the runner-up with $2.3\times$ as many Gaussians. Together, capture-derived learning allowances and training-quality guided density control enable a state-of-the-art quality--size compromise across scene scales without retuning.

Figures & tables

Figure 1: TangoGS overview. The capture sets a learning allowance. Quality feedback controls growth, pruning and the final population retained for continued optimization.

Figure 2: Benchmark quality–size trade-offs. Points are dataset means. Color identifies the scale mechanism and shape the method. Full results: Supplementary § C .

Figure 3: Transfer of released settings. PSNR relative to reference 3DGS at

60

k iterations; The count column gives TangoGS’s count relative to the highest-PSNR method. Speedy-Splat on building (

-4.43

dB) and 3DGS-MCMC are omitted; full results and diagnostics: Supplementary § D .

Figure 4: Capture scale and adaptation ablation. Points show quality changes and count ratios relative to complete TangoGS (star). Complete results: Supplementary § E .

Figure 5: Keep ratio of every capture during training.

Figure 6: Capture statistic ablation. Top: standard and large-capture PSNR versus model size. Bottom: count and LPIPS changes after duplicating every view, held-out evaluation unchanged.

Appendix figures & tables40 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A.1: Selected windows on the handheld captures (Zip-NeRF). Comparators blur thin structure that TangoGS recovers: a door panel and keyhole (NYC), drawer fronts (Alameda), shutter slats (London) and wood grain (Berlin).

Figure A.2: Selected windows on the aerial captures. Vehicles (Rubble), a grass edge (Building), soil texture (Sci-Art) and parking-lot markings (Residence) that comparators smooth over. Reference 3DGS and CDC-GS use up to about twice as many Gaussians as TangoGS on these captures without recovering them.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$N_{G}$	Time
TangoGS	28.21	0.831	0.198	0.97M	194s
LeGS	28.30	0.837	0.183	2.08M	367s
LiteGS	28.07	0.821	0.210	0.99M	168s
3DGS	27.53	0.816	0.211	2.74M	795s
Taming	27.83	0.822	0.204	3.21M	483s
3DGS-MCMC	27.98	0.835	0.185	3.10M	1408s

Appendix

Table C.1: Mip-NeRF 360 (9 scenes). Released baseline settings and the shared TangoGS configuration. Quality and final count are dataset means. Time is training wall-clock on one GH200; the TangoGS prototype includes loading, scoring and diagnostic overhead. A dash denotes unavailable timing.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$N_{G}$	Time
TangoGS	24.77	0.868	0.153	0.79M	208s
LeGS	24.68	0.866	0.155	1.37M	320s
LiteGS	24.85	0.868	0.144	0.99M	196s
3DGS	23.85	0.853	0.168	1.58M	552s
Taming	24.21	0.855	0.169	1.83M	342s
3DGS-MCMC	24.53	0.871	0.146	1.85M	856s

Appendix

Table C.2: Tanks & Temples (2 scenes). Released baseline settings and the shared TangoGS configuration. Quality and final count are dataset means. Time is training wall-clock on one GH200; the TangoGS prototype includes loading, scoring and diagnostic overhead. A dash denotes unavailable timing.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	$N_{G}$	Time
TangoGS	30.55	0.915	0.228	0.84M	115s
LeGS	30.48	0.914	0.229	0.90M	248s
LiteGS	30.34	0.911	0.232	0.99M	158s
3DGS	29.71	0.907	0.239	2.48M	900s
Taming	30.17	0.910	0.234	2.80M	416s
3DGS-MCMC	29.56	0.905	0.233	2.95M	1176s

Appendix

Table C.3: Deep Blending (2 scenes). Released baseline settings and the shared TangoGS configuration. Quality and final count are dataset means. Time is training wall-clock on one GH200; the TangoGS prototype includes loading, scoring and diagnostic overhead. A dash denotes unavailable timing.

Capture	Method	PSNR	SSIM	LPIPS	$N$ (M)	Time (s)
bicycle	LiteGS (1M)	24.84	0.736	0.263	0.99	146
bonsai	LiteGS (1M)	33.17	0.951	0.172	0.99	185
counter	LiteGS (1M)	29.84	0.919	0.170	0.99	194
drjohnson	LiteGS (1M)	29.79	0.909	0.237	0.99	159
flowers	LiteGS (1M)	21.71	0.620	0.315	0.99	176
garden	LiteGS (1M)	27.55	0.857	0.134	0.99	146

Appendix

Table C.4: Released LiteGS on standard scenes. TorchMetrics SSIM; final counts and training time on one GH200. Individual TangoGS results follow in Table C.5 .

Capture	Seed	$N_{G}$ (M)	PSNR	SSIM	LPIPS	Time (s)
bicycle	4114	1.166	25.6566	0.78402	0.21295	203
bicycle	4115	1.170	25.7212	0.78427	0.21289	206
bonsai	4114	0.787	33.0257	0.95151	0.17264	243
bonsai	4115	0.768	33.0668	0.95075	0.17209	243
counter	4114	1.061	29.6930	0.92130	0.16658	250
counter	4115	1.055	29.7756	0.92161	0.16632	239

Appendix

Table C.5: TangoGS: individual standard results. One standard-calibrated configuration on every capture. SSIM uses the reference estimator here; headline cross-method tables retain the TorchMetrics convention for the LiteGS implementation.

	Mip-NeRF 360			Tanks & Temples			Deep Blending
Method	Run	Pub.	$\Delta$	Run	Pub.	$\Delta$	Run	Pub.	$\Delta$
3DGS	27.53	27.21	+0.32	23.85	23.14	+0.71	29.71	29.41	+0.30
Speedy-Splat	26.91	26.94	-0.03	23.37	23.45	-0.08	29.63	29.32	+0.31
Mini-Splatting	27.28	27.34	-0.06	23.15	23.18	-0.03	29.96	29.98	-0.02
Compact 3DGS	27.02	27.08	-0.06	23.38	23.32	+0.06	29.84	29.79	+0.05
LP-3DGS	27.42	27.47	-0.05	23.66	23.60	+0.06	29.61	–	–

Appendix

Table C.6: Reproduced versus published PSNR. Re-run values use released settings;

\Delta

is re-run minus published, with

|\Delta|\geq 0.35

dB in bold. a Both Mip-NeRF 360 means exclude Flowers and Treehill, matching the seven scenes reported by MCMC. b Historical LiteGS v1 ( Liao, 2025 ) , with its own reconstruction and no stated count; context only, without a scored difference. c CDC-GS reports three-seed means and a Mip-NeRF 360 reference baseline

0.26

dB above ours.

	Rubble				Building
Method	PSNR	SSIM	LPIPS	$N$	PSNR	SSIM	LPIPS	$N$
TangoGS	25.63	0.772	0.282	2.96	21.67	0.747	0.271	3.51
3DGS	25.51	0.783	0.268	6.00	21.20	0.744	0.277	7.42
LiteGS (1M)	24.66	0.718	0.349	1.69	21.06	0.698	0.333	1.60
DashGaussian	24.57	0.758	0.292	6.13	19.36	0.662	0.355	4.31
Speedy-Splat	24.20	0.683	0.404	0.58	16.77	0.516	0.531	0.19

Appendix

Table D.1: Rubble and Building. Released-baseline results at

60

k iterations. Counts

N

are in millions; metric conventions follow Appendix B .

	Residence				Sci-Art
Method	PSNR	SSIM	LPIPS	$N$	PSNR	SSIM	LPIPS	$N$
TangoGS	21.06	0.738	0.273	4.57	21.82	0.800	0.230	4.58
3DGS	20.70	0.724	0.293	5.32	20.96	0.784	0.268	2.58
LiteGS (1M)	20.81	0.707	0.316	1.95	21.56	0.785	0.258	2.13
DashGaussian	20.26	0.709	0.304	6.11	20.96	0.780	0.272	3.09
Speedy-Splat	20.16	0.668	0.375	0.53	20.61	0.741	0.345	0.31

Appendix

Table D.2: Residence and Sci-Art. Released-baseline results at

60

k iterations. Counts

N

are in millions; metric conventions follow Appendix B .

	Alameda				London
Method	PSNR	SSIM	LPIPS	$N$	PSNR	SSIM	LPIPS	$N$
TangoGS	23.29	0.772	0.331	3.05	26.96	0.846	0.304	2.80
3DGS	22.99	0.736	0.394	0.87	25.88	0.809	0.376	0.63
LiteGS (1M)	23.00	0.743	0.381	0.98	26.11	0.816	0.360	0.96
DashGaussian	22.89	0.731	0.397	0.99	26.14	0.815	0.366	0.85
Speedy-Splat	21.54	0.671	0.484	0.06	25.00	0.761	0.452	0.06

Appendix

Table D.3: Alameda and London. Released-baseline results at

60

k iterations. Counts

N

are in millions; metric conventions follow Appendix B .

	NYC				Berlin
Method	PSNR	SSIM	LPIPS	$N$	PSNR	SSIM	LPIPS	$N$
TangoGS	28.37	0.872	0.272	1.96	28.38	0.915	0.243	2.71
3DGS	27.12	0.843	0.324	1.06	27.80	0.894	0.294	0.77
LiteGS (1M)	27.73	0.855	0.305	0.98	27.95	0.903	0.275	0.98
DashGaussian	26.81	0.832	0.333	1.03	27.59	0.892	0.296	0.98
Speedy-Splat	25.49	0.797	0.397	0.08	26.58	0.868	0.341	0.08

Appendix

Table D.4: NYC and Berlin. Released-baseline results at

60

k iterations. Counts

N

are in millions; metric conventions follow Appendix B .

Capture	Seed	$N_{G}$ (M)	PSNR	SSIM	LPIPS	Time (s)
alameda	4114	3.050	23.3636	0.77488	0.33054	540
alameda	4115	3.058	23.2152	0.77307	0.33118	557
berlin	4114	2.711	28.4364	0.91594	0.24299	563
berlin	4115	2.713	28.3163	0.91573	0.24307	562
building	4114	3.512	21.7843	0.75075	0.27014	446
building	4115	3.516	21.5522	0.74882	0.27146	450

Appendix

Table D.5: TangoGS: individual large results. One standard-calibrated configuration on every capture. SSIM uses the reference estimator here; headline cross-method tables retain the TorchMetrics convention for the LiteGS implementation.

Capture	Intervention	$N$ (M)	PSNR	SSIM	LPIPS
bicycle	Fixed allowance	0.481	25.227	0.7421	0.2836
	No adaptation	2.504	25.734	0.7937	0.1752
	Both disabled	0.978	25.455	0.7711	0.2245
bonsai	Fixed allowance	0.279	32.325	0.9425	0.1958
	No adaptation	2.790	33.237	0.9533	0.1596
	Both disabled	0.963	33.031	0.9510	0.1718

Appendix

Table E.1: Component controls: standard captures. Interventions defined in Appendix E ; reference SSIM. Default runs are listed in Table C.5 .

Capture	Intervention	$N$ (M)	PSNR	SSIM	LPIPS
alameda	Fixed allowance	0.499	22.902	0.7440	0.3843
	No adaptation	6.153	23.202	0.7669	0.3366
	Both disabled	0.978	22.944	0.7496	0.3722
berlin	Fixed allowance	0.433	27.706	0.8982	0.2847
	No adaptation	6.222	28.246	0.9144	0.2418
	Both disabled	0.963	27.918	0.9040	0.2713

Appendix

Table E.2: Component controls: large captures. Interventions defined in Appendix E ; reference SSIM. Default runs are listed in Table D.5 .

Standard captures
Capture	Shared $N$ (M)	$\Delta$ PSNR	$\Delta$ SSIM	$\Delta$ LPIPS
bicycle	1.168	+0.7927	+0.0396	-0.0424
bonsai	0.778	+0.0508	+0.0012	-0.0034
counter	1.058	-0.0722	+0.0016	-0.0026
drjohnson	0.987	+0.2744	+0.0049	-0.0134
flowers	1.229	+0.2406	+0.0102	-0.0072

Appendix

Table E.3: LiteGS at the final counts selected by TangoGS. All 42 paired counts match exactly. Differences are TangoGS minus native LiteGS, using TorchMetrics SSIM. Positive PSNR/SSIM and negative LPIPS favor TangoGS. Training profiles and populations differ.

Capture	Feedback	$\rho_{T}$	$N$ (M)	PSNR	SSIM	LPIPS
room	Fixed $\rho=0.5$	0.500	0.911	32.6891	0.93355	0.17760
	TangoGS	0.501	0.876	32.6066	0.93323	0.17831
drjohnson	Fixed $\rho=0.5$	0.500	0.998	30.0246	0.91558	0.22447
	TangoGS	0.507	0.987	29.9998	0.91567	0.22461
garden	Fixed $\rho=0.5$	0.500	0.631	27.4832	0.85365	0.14944
	TangoGS	0.536	0.816	27.6288	0.86051	0.13513

Appendix

Table E.4: Fixed versus adaptive keep ratio. Six captures. Only the keep ratio changes;

\rho_{T}

is its terminal value. Reference SSIM.

Regime	Final $N$	Training $N$	Time	PSNR	LPIPS	MB	FPS
	(M)	(M)	(s)	(dB)
Standard	0.925	0.724	183.8	28.04	0.196	229	567
Large	3.268	2.474	514.0	24.65	0.276	811	387

Appendix

Table E.5: Final size and training cost. Equal-capture means. Training population is averaged over optimizer updates; time includes loading, scoring and diagnostics. Quality, saved storage and FPS follow Appendix I .

Capture	$T$ ratio	$N$ ratio	$\Delta$ PSNR	$\Delta$ SSIM	$\Delta$ LPIPS
alameda	0.9136	0.9137	+0.0604	-0.00018	+0.00182
berlin	0.8806	0.8807	+0.0192	-0.00008	+0.00191
bicycle	1.1484	1.1393	+0.0555	+0.00426	-0.00905
bonsai	0.8435	0.8494	-0.0557	-0.00056	+0.00247
building	1.0118	1.0110	+0.0033	+0.00073	+0.00047
counter	0.9862	0.9862	+0.0071	-0.00001	+0.00031

Appendix

Table F.1: Track adjustment versus Total pixels. All 21 captures. Ratios are TangoGS/control; quality changes are TangoGS minus control. Only the metadata allowance changes.

Capture	$T$ ratio	$N$ ratio	$\Delta$ PSNR	$\Delta$ SSIM	$\Delta$ LPIPS
alameda	1.3244	1.3201	+0.0243	+0.00334	-0.00696
berlin	1.4053	1.4006	+0.0588	+0.00246	-0.00611
bicycle	1.3510	1.3265	+0.0775	+0.01028	-0.02063
bonsai	0.7876	0.7951	-0.0905	-0.00091	+0.00338
building	0.6852	0.6934	-0.2278	-0.01223	+0.01589
counter	0.9606	0.9626	+0.0282	+0.00027	+0.00026

Appendix

Table F.2: Track adjustment versus SfM points. All 21 captures. Ratios are TangoGS/control; quality changes are TangoGS minus control. Only the metadata allowance changes.

Capture	Statistic	$N$ orig.	$N$ dup.	$\Delta$ PSNR	$\Delta$ SSIM	$\Delta$ LPIPS
bicycle	Track-adjusted	1.168	1.207	+0.036	+0.0012	-0.0020
bicycle	Total pixels	1.025	1.471	+0.089	+0.0108	-0.0231
bonsai	Track-adjusted	0.778	0.857	-0.024	-0.0001	-0.0013
bonsai	Total pixels	0.916	1.408	+0.190	+0.0015	-0.0051
counter	Track-adjusted	1.058	1.022	-0.063	-0.0007	-0.0002
counter	Total pixels	1.073	1.447	+0.038	+0.0004	-0.0045

Appendix

Table F.3: Exact view duplication. Counts are in millions, and quality changes are duplicated minus original (reference SSIM). Allowance ratios are

1

for track-adjusted pixels/SfM points and

\sqrt{2}

for total pixels. Final rows are equal-capture means.

Method	Controls to revisit	Calibration issue
Fixed-threshold
3DGS ( Kerbl et al., 2023 )	Growth gradient threshold	Threshold has no direct count or quality interpretation.
Speedy-Splat ( Hanson et al., 2025 )	Growth threshold; pruning fraction	Retention depends on the population admitted by growth.
Preset-budget
Taming 3DGS ( Mallick et al., 2024 )	SfM multiplier; growth threshold	A larger budget need not admit more growth candidates.
Mini-Splatting ( Fang and Wang, 2024 )	Population limits; sampling fraction	Capacity before simplification and final retention both need selection.

Appendix

Table G.1: Candidate retuning controls for large captures. Entries describe the evaluated implementations and retain the scale-mechanism groups used in the main paper. They identify controls to investigate, not measured tuning effort or guaranteed remedies.

Capture	Method	PSNR	SSIM	LPIPS	$N$ (M)
Mip-NeRF 360	Taming 3DGS	27.83	0.822	0.204	3.21
Mip-NeRF 360	TangoGS on Taming 3DGS	27.61	0.812	0.216	0.96
Tanks & Temples	Taming 3DGS	24.21	0.855	0.169	1.83
Tanks & Temples	TangoGS on Taming 3DGS	24.49	0.863	0.166	0.78
Deep Blending	Taming 3DGS	30.17	0.910	0.234	2.80
Deep Blending	TangoGS on Taming 3DGS	30.08	0.912	0.242	0.78

Appendix

Table H.1: The controller on reference 3DGS. Released Taming 3DGS and the same controller run on its code, both under the reference evaluation script. Standard scenes are dataset means over the two-seed mean per scene; large captures use the common

60

k iterations and one seed. Counts in millions.

Method	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$
	Standard (13)				Large (8)
TangoGS	28.04	0.196	229	567	24.65	0.276	811	387
LeGS	28.08	0.186	445	520	23.41	0.348	233	881
LiteGS	27.92	0.203	246	538	24.11	0.322	350	601
3DGS	27.30	0.209	625	203	24.02	0.312	764	216
Taming	27.63	0.203	728	220	23.77	0.346	356	329

Appendix

Table I.1: Quality, storage and rendering throughput. Equal-capture means; two seeds per capture for TangoGS. PSNR is in dB; MB measures saved model size. Taming uses its released standard budgets and

15\times

on large captures. The last row gives Compact 3DGS's separate encoded-payload estimate, not saved-file bytes.

Method	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$
	Bicycle				Flowers
TangoGS	25.69	0.213	290	608	21.99	0.296	305	436
LeGS	25.73	0.167	840	446	22.04	0.282	651	477
LiteGS	24.84	0.263	246	673	21.71	0.315	246	476
3DGS	25.20	0.208	1231	116	21.57	0.335	718	242
Taming	25.50	0.195	1485	144	21.82	0.334	897	220

Appendix

Table I.2: Per-capture quality, storage and throughput. PSNR (dB) and VGG LPIPS evaluate the same saved checkpoints as MB and FPS. The last row is Compact 3DGS's encoded-payload estimate. TangoGS entries average both seeds.

Method	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$
	Garden				Stump
TangoGS	27.63	0.135	202	795	27.46	0.199	212	676
LeGS	28.09	0.089	836	400	27.40	0.177	569	505
LiteGS	27.55	0.134	246	786	27.13	0.214	247	612
3DGS	27.49	0.106	1026	162	26.64	0.215	1063	200
Taming	27.78	0.100	1273	191	26.97	0.201	1207	209

Appendix

Table I.3: Per-capture quality, storage and throughput. PSNR (dB) and VGG LPIPS evaluate the same saved checkpoints as MB and FPS. The last row is Compact 3DGS's encoded-payload estimate. TangoGS entries average both seeds.

Method	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$
	Treehill				Room
TangoGS	23.31	0.313	267	630	32.61	0.178	217	473
LeGS	23.00	0.303	629	529	32.62	0.181	259	577
LiteGS	23.37	0.323	247	600	32.63	0.183	245	516
3DGS	22.51	0.324	824	203	31.54	0.205	323	198
Taming	23.06	0.313	935	201	32.05	0.199	384	206

Appendix

Table I.4: Per-capture quality, storage and throughput. PSNR (dB) and VGG LPIPS evaluate the same saved checkpoints as MB and FPS. The last row is Compact 3DGS's encoded-payload estimate. TangoGS entries average both seeds.

Method	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$
	Counter				Kitchen
TangoGS	29.73	0.166	262	490	32.39	0.112	222	554
LeGS	29.82	0.170	207	493	32.58	0.112	354	387
LiteGS	29.84	0.170	246	459	32.37	0.114	246	521
3DGS	29.07	0.191	268	208	31.31	0.122	392	178
Taming	29.04	0.190	295	218	31.83	0.118	401	202

Appendix

Table I.5: Per-capture quality, storage and throughput. PSNR (dB) and VGG LPIPS evaluate the same saved checkpoints as MB and FPS. The last row is Compact 3DGS's encoded-payload estimate. TangoGS entries average both seeds.

Method	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$
	Bonsai
TangoGS	33.05	0.172	193	448
LeGS	33.41	0.164	308	508
LiteGS	33.17	0.172	246	441
3DGS	32.43	0.190	264	288
Taming	32.43	0.188	294	246

Appendix

Table I.6: Per-capture quality, storage and throughput. PSNR (dB) and VGG LPIPS evaluate the same saved checkpoints as MB and FPS. The last row is Compact 3DGS's encoded-payload estimate. TangoGS entries average both seeds.

Method	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$
	Truck				Train
TangoGS	26.47	0.139	166	594	23.06	0.168	227	414
LeGS	26.58	0.128	472	449	22.78	0.182	209	491
LiteGS	26.57	0.113	246	483	23.14	0.174	246	382
3DGS	25.46	0.142	512	231	22.24	0.195	272	225
Taming	25.91	0.129	641	245	22.50	0.208	269	308

Appendix

Table I.7: Per-capture quality, storage and throughput. PSNR (dB) and VGG LPIPS evaluate the same saved checkpoints as MB and FPS. The last row is Compact 3DGS's encoded-payload estimate. TangoGS entries average both seeds.

Method	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$
	Dr Johnson				Playroom
TangoGS	30.00	0.225	245	566	31.10	0.232	172	692
LeGS	29.96	0.232	244	713	30.99	0.225	202	790
LiteGS	29.79	0.237	246	522	30.88	0.226	246	529
3DGS	29.35	0.236	775	159	30.07	0.241	456	235
Taming	29.78	0.234	812	214	30.57	0.234	577	251

Appendix

Table I.8: Per-capture quality, storage and throughput. PSNR (dB) and VGG LPIPS evaluate the same saved checkpoints as MB and FPS. The last row is Compact 3DGS's encoded-payload estimate. TangoGS entries average both seeds.

Method	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$
	Rubble				Building
TangoGS	25.63	0.282	733	441	21.67	0.271	872	443
LeGS	24.36	0.355	374	765	20.08	0.376	228	953
LiteGS	24.66	0.349	420	666	21.06	0.333	398	555
3DGS	25.51	0.268	1488	152	21.20	0.277	1841	108
Taming ( $15\times$ )	25.18	0.334	716	294	20.69	0.348	524	246

Appendix

Table I.9: Per-capture quality, storage and throughput. PSNR (dB) and VGG LPIPS evaluate the same saved checkpoints as MB and FPS. The last row is Compact 3DGS's encoded-payload estimate. TangoGS entries average both seeds.

Method	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$
	Residence				Sci-Art
TangoGS	21.06	0.273	1133	381	21.82	0.230	1136	309
LeGS	20.27	0.328	444	638	20.50	0.289	367	736
LiteGS	20.81	0.316	484	672	21.56	0.258	529	575
3DGS	20.70	0.293	1320	131	20.96	0.268	640	195
Taming ( $15\times$ )	20.59	0.315	727	232	21.83	0.267	507	249

Appendix

Table I.10: Per-capture quality, storage and throughput. PSNR (dB) and VGG LPIPS evaluate the same saved checkpoints as MB and FPS. The last row is Compact 3DGS's encoded-payload estimate. TangoGS entries average both seeds.

Method	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$
	Alameda				London
TangoGS	23.29	0.331	757	364	26.96	0.304	694	448
LeGS	21.98	0.420	124	981	26.18	0.365	144	993
LiteGS	23.00	0.381	242	581	26.11	0.360	237	720
3DGS	22.99	0.394	215	283	25.88	0.376	155	362
Taming ( $2\times$ )	22.87	0.421	108	379	25.44	0.407	74	473

Appendix

Table I.11: Per-capture quality, storage and throughput. PSNR (dB) and VGG LPIPS evaluate the same saved checkpoints as MB and FPS. The last row is Compact 3DGS's encoded-payload estimate. TangoGS entries average both seeds.

Method	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	MB $\downarrow$	FPS $\uparrow$
	NYC				Berlin
TangoGS	28.37	0.272	487	390	28.38	0.243	672	320
LeGS	26.76	0.343	93	1028	27.16	0.306	91	957
LiteGS	27.73	0.305	244	515	27.95	0.275	242	526
3DGS	27.12	0.324	263	260	27.80	0.294	190	237
Taming ( $2\times$ )	26.32	0.363	93	438	27.45	0.313	103	324

Appendix

Table I.12: Per-capture quality, storage and throughput. PSNR (dB) and VGG LPIPS evaluate the same saved checkpoints as MB and FPS. The last row is Compact 3DGS's encoded-payload estimate. TangoGS entries average both seeds.

Capture	$N_{\rm auto}$	$N_{r}$	$N^{*}$	Excess	Qualifies
Bicycle	1.166	2.233	1.709	–	no
Bonsai	0.787	1.531	0.591	33.2%	yes
Rubble	2.960	5.765	4.375	–	no
Residence	4.568	8.991	4.568	0.0%	yes
London	2.798	5.588	2.798	0.0%	yes
Berlin	2.711	5.413	2.046	32.5%	yes

Appendix

Table J.1: Capacity selection with seed 4114. Counts are millions of Gaussians;

N_{r}

is the

2\times

-allowance reference and

N^{*}

is the smallest evaluated model satisfying both quality limits. Excess count is reported only for qualifying automatic models.

Capture	PSNR loss (dB)	SSIM loss	LPIPS loss
Bicycle	0.160	0.0151	0.0356
Bonsai	0.175	0.0015	0.0088
Rubble	0.408	0.0253	0.0364
Residence	0.248	0.0128	0.0161
London	-0.019	0.0035	0.0149
Berlin	-0.039	0.0022	0.0101

Appendix

Table J.2: Automatic-model quality losses relative to the

2\times

-allowance reference. Negative values favor the automatic model. SSIM is descriptive; qualification uses PSNR and LPIPS. The joint limits are

0.30

dB PSNR and

0.020

VGG LPIPS.

Figure J.1: Quality–size objective with one seed. The vertical coordinate is the larger of the PSNR and LPIPS losses divided by their respective limits; values at most one satisfy both limits. Stars mark TangoGS’s automatic point and open circles the smallest evaluated qualifying model. Lines connect evaluated endpoints.

Figure J.2: Quality–count curves for Bicycle, Bonsai and Rubble. Stars mark the automatic operating points, squares the

2\times

-allowance references and dotted lines the PSNR and LPIPS limits relative to those references.

Figure J.3: Quality–count curves for Residence, London and Berlin, with the same conventions as Fig. J.2 .

3D Gaussian Splatting

3D Gaussian Splatting represents scenes as optimizable 3D Gaussian primitives for fast, photorealistic rendering from new viewpoints.

See all 986 papers

ZipSplat: Fewer Gaussians, Better Splats
Jun 3, 2026Alexander Veicht, Sunghwan Hong, Dániel Baráth +13D Gaussian Splatting Multi-View 3D Reconstruction
ETH Zürich · Microsoft
Prune Wisely, Reconstruct Sharply: Compact 3D Gaussian Splatting via Adaptive Pruning and Difference-of-Gaussian Primitives
Feb 27, 2026Haoran Wang, Guoxi Huang, Fan Zhang +23D Gaussian Splatting 3D Gaussian Pruning
School of Computer Science, University of Bristol, Bristol, UK
TideGS: Scalable Training of Over One Billion 3D Gaussian Splatting Primitives via Out-of-Core Optimization
May 19, 2026Chonghao Zhong, Linfeng Shi, Hua Chen +43D Gaussian Splatting 3D Scene Representation
Hong Kong University of Science and Technology · Great Wall Motor · Tsinghua University +1

Jun 3, 2026J/K move · Enter open · S save

ZipSplat: Fewer Gaussians, Better Splats

Alexander Veicht, Sunghwan Hong, Dániel Baráth, Marc Pollefeys

ETH Zürich · Microsoft

3D Gaussian Splatting Multi-View 3D Reconstruction Feed-Forward 3D Reconstruction

Open paper AskPDF arXiv

Feed-forward 3D Gaussian Splatting reconstructs scenes from images in seconds. Pixel-aligned methods keep every Gaussian grounded in an observation, but produce redundant, potentially misaligned Gaussians across views. Token-based methods instead aggregate all views into a fixed set of learned tokens but lack explicit observation support and use a fixed reconstruction budget. We introduce ZipSplat, a method that predicts a compact, observation-grounded scene representation whose size is set at inference. It derives each scene token from an image patch rather than a learned query, and decodes it into freely placed 3D Gaussians. Tokens across views are merged into a compact set whose size is adjustable at inference. However, without geometric supervision, decoded Gaussians can drift from their source patch, and merging can mix tokens from different surfaces. We therefore introduce two geometric losses to ground and shape this representation. A cone loss ties each token's Gaussians to its source patch, and GeoCE aligns token similarity with 3D proximity so that tokens observing the same surface merge together. Building on these observation--geometry associations, we further introduce TokenBA, a token-level bundle adjustment that refines scene tokens and camera poses in a few seconds. ZipSplat requires neither camera poses nor intrinsics, yet outperforms pixel-aligned and token-based baselines on DL3DV and RealEstate10K, even those given ground-truth cameras. It further generalizes zero-shot to Mip-NeRF360 and ScanNet++. Our project page is at https://veichta.com/zipsplat.

Abstract

3D Gaussian Splatting

Explore similar work

ZipSplat: Fewer Gaussians, Better Splats

Prune Wisely, Reconstruct Sharply: Compact 3D Gaussian Splatting via Adaptive Pruning and Difference-of-Gaussian Primitives

TideGS: Scalable Training of Over One Billion 3D Gaussian Splatting Primitives via Out-of-Core Optimization